Prediction of current and future gully erosion susceptibility in the Oued Zat watershed in the High Atlas region of Morocco using an SVM-based machine learning approach

Abstract

Gully erosion constitutes a major natural hazard causing significant land degradation. Here, we aimed to assess the current and future susceptibility of gully erosion in the Oued Zat watershed (High Atlas, Morocco) using machine learning (ML). Leveraging 210 gully and 400 non-gully points along with 16 conditioning factors from precipitation, remote sensing, geological, and soil data, we employed five support vector machine (SVM) models based on linear, radial, and polynomial kernels for gully susceptibility modeling. Performance evaluation included specificity, recall, receiver operating characteristic area under the curve (ROC–AUC), precision, and F1 score metrics. Future vulnerability in the 21st century was assessed under two shared socioeconomic pathways (SSP2-4.5 and SSP5-8.5) climate scenarios. Our results demonstrated superior performance by SVM-Poly (recall: 0.93; AUC–ROC: 0.92; precision: 0.89; F1-Score: 0.91) followed by SVM-RBF (recall: 0.84; AUC–ROC: 0.86; precision: 0.78; F1-Score: 0.81), while linear models were the least accurate. Current predictions showed a prevalence of high- and very high-risk areas to gullying, together accounting for over 38% of the watershed, particularly in the south, while low- and very low-risk areas (26%) dominated in the north. Future projections anticipate declining very low, low, and moderate-risk areas. Initially, high- and very high-risk zones will increase in extent, followed by a gradual decline in very high-risk areas. In contrast, high-risk areas will continue increasing, reaching around 20% and 40% by 2100 under SSP2-4.5 and SSP5-8.5, respectively. Our findings not only reveal vulnerability to erosion but also demonstrate a practical, effective approach for decision-making, especially considering the impacts of climate change in the region.